Modelpublic

ZeyuLing/hftrainer-motionlab-humanml3d

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

0likes

Model Card

MotionLab - Unified Human Motion Generation and Editing

Text-to-motion baseline integrated into the hftrainer Model Zoo. The runtime is self-contained under hftrainer.models.motion.motionlab.network and does not import the original repository at inference time.


Task	Text-to-Motion (T2M), motion generation / editing research stack
Bundle / Pipeline	`MotionLabBundle` / `MotionLabPipeline`
Processed HF artifact	`ZeyuLing/hftrainer-motionlab-humanml3d`
Motion representation	HumanML3D-263 (263-dim, 20 fps, 22 joints)
Architecture	RFMotion / MotionFlow Transformer with CLIP text conditioning
Paper	MotionLab: Unified Human Motion Generation and Editing via the Motion-Condition-Motion Paradigm, Guo et al., ICCV 2025 - arXiv:2502.02358
Original code	https://github.com/Diouo/MotionLab

Weights

Self-contained hftrainer artifact:

Artifact	Location	Contents	Status
MotionLab HumanML3D	`ZeyuLing/hftrainer-motionlab-humanml3d`	`motionflow.ckpt` + `configs/` + `Mean.npy` / `Std.npy` + `mean_motion.npy` / `std_motion.npy` + `model_index.json`	public Hub artifact
local mirror	`checkpoints/baselines/motionlab`	same layout	optional local cache

Use directly from the Hub:

python

from hftrainer.pipelines.motionlab import MotionLabPipeline

pipe = MotionLabPipeline.from_pretrained(
    "ZeyuLing/hftrainer-motionlab-humanml3d",
    device="cuda",
)
motions = pipe.infer_t2m(
    ["a person walks forward then sits down"],
    [120],
)  # list of (T, 263)

For a local mirror:

python

pipe = MotionLabPipeline.from_pretrained("checkpoints/baselines/motionlab", device="cuda")

Motion Representation

MotionLab natively generates HumanML3D-263 at 20 fps. For shared SMPL and MotionStreamer-272 evaluation, use the validated bridge:

text

HumanML3D-263 -> SMPL motion_135 via IK refine-80 -> MotionStreamer-272

The artifact contains both the HumanML3D denormalization statistics and MotionLab's internal motion statistics so the published pipeline does not depend on a separate dataset checkout.

HumanML3D Leaderboard Metrics

The row below uses the shared HumanML3D official-test caption protocol and the HML263 round-trip GT reference for SMPL-based evaluators.

Evaluator	R1 up	R2 up	R3 up	FID down	MM down	Div up
MotionStreamer-272	0.6367	0.7882	0.8529	25.4469	17.9756	25.5355
MotionCLIP-135 no-L2	0.4807	0.6457	0.7353	102.7770	41.5472	23.0179

Physical metrics:

Slide down	Float down	Jitter down	Dynamic down
2.4231	4.0795	5.8493	24.3519

Implementation Notes

—Artifact inference imports only hftrainer.models.motion.motionlab.network.
—Config targets are rewritten from the original rfmotion.* namespace into the vendored hftrainer namespace before model construction.
—The default inference stage is demo, matching the validated qualitative HumanML3D T2M setting.